MULTI-STEP NEURAL NETWORK BASED SYSTEM IDENTIFICATION AND NMPC FOR AN UNDERACTUATED UNMANNED SURFACE VEHICLE

Date

2026-1-22

Author

Yıldırım, Mustafa

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

39
views

0
downloads

Cite This

This thesis focuses on learning a dynamics model for an underactuated 3-DOF unmanned surface vehicle (USV) using neural networks and deploying the learned model inside nonlinear model predictive control (NMPC). Physics-based maneuvering models provide useful structure, but accuracy can be affected by parameter uncertainty, noisy measurements, and actuator behavior. In real operation, the USV dynamics depend on current velocity and the actual thrust, while the available signals are noisy velocity measurements and thruster commands that differ from delivered thrust because of nonlinearities and delays. To reduce this mismatch, the neural identifiers use a short history of body-frame velocities together with a short history of thruster commands, enabling the networks to implicitly estimate effective thrust and handle noise. Three predictors are emphasized and compared: (i) a conventional 3-DOF Fossen maneuvering model, (ii) a multi-step MLP that outputs the full prediction horizon in a single forward pass, and (iii) a sequence-to-sequence (encoder-decoder) LSTM with the same horizon-length output. The neural models are trained on trajectory segments whose length matches the NMPC horizon to limit drift in multi-step predictions. Data are generated and the controllers are tested in a Gazebo simulation environment using a Clearpath Robotics Heron USV model. The NMPC formulation is designed to accept either the learned model or the physics-based model and is evaluated on bodyframe velocity tracking and position/trajectory tracking. Results indicate that multistep neural predictors can provide accurate horizon predictions and support real-time NMPC.

Subject Keywords

System Identification, Model Predictive Control, Optimization, Neural Networks

URI

https://hdl.handle.net/11511/118387

Collections

Graduate School of Natural and Applied Sciences, Thesis